AI Datacenter

An AI datacenter (also written as AI data center) is a facility designed to run artificial intelligence workloads on dense groups of GPUs or other accelerators. It combines compute systems, high-bandwidth networks, storage, electrical equipment, and cooling appropriate to the installed hardware. AI datacenters train and serve large language models and other models, but the term does not imply one fixed rack design or one mandatory cooling method. The International Energy Agency (IEA) estimated that all data centers used about 415 terawatt-hours (TWh), or 1.5% of global electricity, in 2024 and projected about 945 TWh in 2030, with AI as the largest driver of the increase ^[1].

Unlike a general-purpose data center, an AI-focused facility is usually optimized for repeated matrix operations, rapid communication among accelerators, and the movement of large model and dataset states. Those requirements can produce unusually high rack power, network traffic, and heat loads. The result is an engineering system in which chip choice, cluster topology, power delivery, cooling, storage, and software must be designed together rather than treated as independent purchases ^[7][8][9].

What is an AI datacenter?

An AI datacenter is defined by its workload and systems architecture, not simply by the presence of an accelerator. Training a large model can require many devices to exchange gradients and other intermediate data repeatedly. Serving, or inference, can instead emphasize response time, throughput, availability, and placement near users. A facility may support both patterns, along with model fine-tuning, data preparation, evaluation, and scientific computing ^[9][12].

The physical scale varies widely. An enterprise deployment may consist of a small number of accelerator servers in an existing computer room. A cloud deployment may span many halls or multiple sites. The IEA uses a household comparison to convey the upper end: it estimated that a typical AI-focused data center consumes as much electricity as 100,000 households, while the largest facilities then under construction could consume 20 times as much ^[1]. This is an illustrative comparison, not a standard size requirement.

AI equipment also does not establish a universal cooling threshold. NVIDIA documents an eight-GPU DGX B300 system at 14.5 kW nominal power and 15 kW system maximum. Its deployment guide describes two such systems in a rack with air cooling at about 30 kW average and four systems with air cooling plus active rear-door heat exchangers at about 58 kW average. By contrast, a GB300 NVL72 rack is liquid cooled and can require up to 142 kW ^[7][8][9]. These examples show why cooling has to be specified for the actual platform and rack layout.

What is the AI datacenter buildout?

The current buildout includes new campuses, expansions of existing cloud regions, leased colocation space, utility interconnections, network backbones, and accelerator purchases. The IEA estimated that global investment in data centers reached about $500 billion in 2024, nearly twice its 2022 level. It also estimated that about 20% of planned data-center projects could face delays if grid constraints are not addressed, noting that new transmission lines can take four to eight years in advanced economies ^[1].

The phrase "AI datacenter buildout" therefore covers more than construction. A company can increase available compute by installing more servers in an existing building, leasing powered space, contracting with a specialized cloud provider, commissioning a new campus, or combining these approaches. Reported capital expenditure may also include non-AI assets, while leased capacity may appear differently in financial statements. Comparisons need to retain those distinctions.

Hyperscaler capital expenditure

Major technology companies have reported very large infrastructure budgets for 2026. The figures below are the latest primary disclosures located for this review, but they are not a comparable measure of AI-only spending. They use different fiscal periods and definitions, and most cover company-wide assets beyond AI datacenters.

Scroll sideways for more →

Adding these numbers would produce a large total, but it would not produce an audited total for AI datacenters. Amazon's figure covers all of Amazon, Alphabet includes Other Bets, Microsoft's number includes finance leases, Meta explicitly includes finance-lease principal payments, and Oracle reports an already completed fiscal year ^{[2][3][4][5][6]}. A careful comparison reports the basis beside each number instead of assigning an unsupported AI percentage.

Microsoft's $627 billion commercial remaining performance obligation in fiscal 2026 Q3 is also not an Azure-only backlog or a direct measure of unpowered datacenter demand. The company said that amount included OpenAI, had a weighted average duration of about two and a half years, and covered commercial obligations. In the same call, Microsoft separately said it expected compute capacity to remain constrained through at least 2026 ^[4]. Those statements can inform infrastructure demand without being combined into an $80 billion "power-constrained Azure backlog."

How is an AI datacenter built? (Hardware architecture)

An AI facility contains conventional datacenter systems as well as workload-specific equipment. Utility feeds and substations deliver power to switchgear, transformers, uninterruptible power supplies, and distribution equipment. Racks hold compute, network, and storage systems. Cooling equipment removes the heat, while control systems monitor electrical, mechanical, network, and security conditions. Redundancy choices determine how failures are isolated and how much spare capacity is maintained.

For AI, the design process often starts with the accelerator platform and intended cluster size. Those choices determine server power, rack weight, network interface count, cable reach, cooling media, storage throughput, and the number of racks that can share a power or cooling zone. NVIDIA's DGX B300 planning guide, for example, treats power, cooling, and space as interrelated constraints and supplies separate high-density and lower-density deployment patterns ^[8].

GPU clusters

The main compute layer is commonly a GPU cluster, although custom accelerators are also used. A cluster joins servers into a scheduled resource for training, inference, or high-performance computing. The useful scale is not just the number of chips: accelerator memory, interconnect topology, network oversubscription, storage behavior, software libraries, job scheduling, and failure recovery all affect completed workload time.

NVIDIA sells systems at more than one physical scale. An NVIDIA DGX B300 server contains eight B300 GPUs, 2.3 TB of aggregate GPU memory, and eight 800 Gb/s cluster-network connections. Its user guide lists 14.5 kW power consumption and a 15 kW system maximum ^[7]. The company's 2026 planning guide shows that operators can deploy two air-cooled systems per rack or four air-cooled systems with active rear-door heat exchangers, illustrating that a Blackwell-generation server is not automatically a direct-liquid-cooled rack ^[8].

The GB300 NVL72 is a different design. It integrates 72 Blackwell Ultra GPUs and 36 Grace CPUs through fifth-generation NVLink. NVIDIA specifies a liquid-cooled rack, 130 TB/s of aggregate NVLink bandwidth, and up to 142 kW for the full rack. Each GPU has up to 1,800 GB/s of NVLink bandwidth, and the rack's compute trays use 800 Gb/s ConnectX-8 interfaces for scale-out communication ^[9]. These are vendor specifications for that product, not a generic performance or power number for every AI rack.

Other operators use or offer non-NVIDIA accelerators. Alphabet's 2025 Q4 call said its infrastructure includes NVIDIA GPUs and its own Tensor Processing Units. Amazon's 2025 results said Trainium2 was fully subscribed, with 1.4 million chips landed, and that the more than 500,000-chip Project Rainier cluster was being used by Anthropic ^[2][3]. These first-party disclosures demonstrate a mixed accelerator market without establishing a complete market-share ranking.

How are AI datacenters networked? (NVLink and InfiniBand)

AI networks operate at several scopes. A local interconnect joins accelerators inside a server or rack. A scale-out fabric joins servers and racks for distributed jobs. Separate networks may carry storage, management, tenant, and internet traffic. The design objective is not only high peak bandwidth; distributed training also depends on predictable collective communication, congestion control, fault isolation, and a topology that matches the job placement strategy.

Within NVIDIA rack-scale systems, NVLink and NVSwitch form an accelerator fabric. In the GB300 NVL72, nine NVSwitch trays connect all 72 GPUs, creating 130 TB/s of aggregate rack bandwidth. NVIDIA describes 1,800 GB/s per GPU for fifth-generation NVLink ^[9]. These numbers describe the in-rack fabric and should not be confused with the separate network used to join racks.

At scale-out, both InfiniBand and Ethernet-based designs are available. NVIDIA's Quantum-X800 InfiniBand switch provides 144 ports at 800 Gb/s per port and includes adaptive routing, telemetry-based congestion control, and in-network computing features ^[10]. Ethernet implementations can use RDMA and congestion-management mechanisms as well. The Ultra Ethernet Consortium released Specification 1.0 in June 2025 as an Ethernet-based communications stack for AI and high-performance computing, covering network adapters, switches, optics, cables, transport, and security ^[11].

Public product documentation supports the existence and capabilities of both approaches, but it does not support a universal claim that one held a majority of all new deployments. Network choice depends on the accelerator platform, existing operational tooling, supplier ecosystem, desired interoperability, topology, workload, and cost. Product bandwidth also does not by itself establish application-level latency or training performance.

How are AI datacenters cooled?

Cooling removes heat from the IT equipment and rejects it to the outside environment. Designs may use room air, contained aisles, rear-door heat exchangers, direct liquid cold plates, immersion systems, or combinations of those methods. The heat-rejection stage can use dry coolers, chillers, cooling towers, or other equipment. Because "liquid cooling" can refer to different parts of this chain, it is useful to distinguish liquid delivered to a chip cold plate from water used in a facility cooling tower.

Air cooling remains practical for some current accelerator systems. NVIDIA's DGX B300 guide describes two systems per rack at about 30 kW average with unassisted air cooling and four systems per rack at about 58 kW average with active rear-door heat exchangers. The four-system pattern has a stated peak of 76 kW ^[8]. This does not mean every room can accept that heat load without modification; it means the vendor documents an air-cooled system-level path.

Direct liquid cooling places a liquid loop close to processors through cold plates and coolant distribution units. It can support higher heat flux and reduce the air volume needed around a rack. The GB300 NVL72 uses this approach, with integrated leak detection and up to 142 kW of rack power ^[9]. CoreWeave also says its infrastructure uses closed-loop liquid cooling where appropriate, but its filing describes a portfolio of facilities rather than a claim that every deployment uses the same method ^[12].

Cooling and water impacts are not interchangeable. A closed-loop liquid circuit can recirculate its working fluid while the facility's heat-rejection equipment still uses electricity or water. Conversely, an air-cooled server can be installed in a building that uses evaporative cooling. Measuring direct water use, indirect water associated with electricity generation, and total facility energy is necessary before comparing designs ^[20][21].

How much power does an AI datacenter use?

There is no single power number for an AI datacenter. A server may draw kilowatts, a rack tens or more than one hundred kilowatts, and a campus hundreds of megawatts or several gigawatts when fully built. Nameplate, contracted, planned, secured, active, and average operating power are different measurements. Announcements often state a future campus ceiling rather than current consumption.

Scale of power consumption

The following examples retain the status and date supplied by the source. They should not be read as directly comparable operating loads.

Scroll sideways for more →

Grid access can be as important as server procurement. The IEA estimated that around 20% of planned data-center projects could be at risk of delay without action on grid constraints. It cited four-to-eight-year transmission construction timelines in advanced economies and longer waits for transformers and cables ^[1]. Microsoft separately said in April 2026 that it expected its compute capacity to remain constrained at least through the end of 2026 despite efforts to bring GPU, CPU, and storage capacity online faster ^[4].

Global totals also require careful scope. The IEA's 415 TWh estimate for 2024 and 945 TWh projection for 2030 cover all data centers, not AI alone. It identified AI as the most important source of growth alongside other digital services. The IEA projected that the United States would account for the largest increase, followed by China, and that data centers would account for nearly half of US electricity-demand growth through 2030 ^[1].

Notable facilities

SpaceXAI Colossus (Memphis, Tennessee)

Colossus is the AI computing site built by xAI, now described on first-party pages under SpaceXAI, in Memphis. Its public project page says the first system was built in 122 days and then doubled to 200,000 GPUs in another 92 days. That page contains both 180,000 and 200,000 H100 labels in different sections, so the safest current equipment statement comes from the later May 2026 partnership announcement ^[14][15].

That later announcement says Colossus 1 contains more than 220,000 NVIDIA GPUs, including H100, H200, and GB200 accelerators, and supports training, fine-tuning, inference, and high-performance computing. It announced that Anthropic would use additional compute for its Claude subscription services ^[14]. The source does not provide a current site power total or the earlier article's exact component split, so neither is asserted here.

Stargate Project

The Stargate Project was announced by OpenAI and SoftBank on January 21, 2025 as a new company intending to invest $500 billion over four years in US AI infrastructure, beginning with $100 billion. The announcement names SoftBank, OpenAI, Oracle, and MGX as initial equity funders, with SoftBank holding financial responsibility and OpenAI operational responsibility. It did not publish equity percentages ^[16].

Stargate's disclosed scale changed as agreements and sites developed. In September 2025, OpenAI said five newly announced sites, the flagship Abilene site, and ongoing CoreWeave projects brought the platform to nearly 7 GW of planned capacity and more than $400 billion of investment over the following three years. The five sites were in Shackelford County, Texas; Doña Ana County, New Mexico; Wisconsin; Lordstown, Ohio; and Milam County, Texas, with a potential 600 MW expansion near Abilene also described ^[17].

In April 2026, OpenAI said it had surpassed its goal of securing 10 GW of US AI infrastructure, with more than 3 GW added during the preceding 90 days. "Secured" is the company's term and should not be read as "operational." OpenAI described Abilene as the flagship site and said it used NVIDIA GB200 systems on Oracle Cloud Infrastructure. It also said Abilene's cooling loop recirculates water through sealed pipes after an initial fill, rather than using traditional evaporative cooling towers ^[18].

CoreWeave infrastructure

CoreWeave operates a distributed cloud platform built around accelerated computing. Its 2025 Form 10-K reports 43 data centers and more than 850 MW of active power as of December 31, 2025, up from 32 centers and about 360 MW a year earlier. It also reports about 3.1 GW of contracted power for future deployment and $60.7 billion of remaining performance obligations ^[12]. These portfolio measures supersede the older 32-center figure in the prior article.

The same filing says CoreWeave's facilities were in six countries across the United States, Europe, and Canada and ranged from smaller inference-oriented sites to larger training sites. It describes closed-loop liquid cooling and says liquid cooling is incorporated where appropriate. CoreWeave also disclosed that much of its existing space was leased or licensed from third-party datacenter providers, while it was beginning to develop some facilities itself ^[12].

CoreWeave and Meta announced in an April 2026 SEC exhibit an expanded agreement worth approximately $21 billion through December 2032. The capacity was to be distributed across multiple locations and include early NVIDIA Vera Rubin deployments ^[13]. This is a customer agreement, not an equipment inventory or a statement that all contracted capacity was operating when announced.

What is the environmental impact of AI datacenters?

Environmental effects occur at the facility, in the electricity system, and through equipment supply chains. Important measures include electricity use, greenhouse-gas intensity, direct and indirect water consumption, land and transmission requirements, construction materials, and hardware replacement. Results depend strongly on location, grid mix, cooling design, utilization, and the boundaries chosen for an assessment ^[1][20][21].

Electricity consumption and carbon

The IEA estimated global data-center electricity use at 415 TWh in 2024 and projected about 945 TWh in 2030. In 2024, the United States accounted for 45% of the total, China 25%, and Europe 15%. The agency's base case projects emissions from datacenter electricity use rising from about 180 million tonnes to 300 million tonnes by 2035, while noting substantial uncertainty in both AI uptake and energy-system development ^[1].

For the United States, Lawrence Berkeley National Laboratory's June 2026 update projects a 2030 reference case of 649 TWh, equal to 11.8% of US electricity. Its compounded uncertainty scenarios span 521 to 843 TWh, or 9.5% to 15.3% of US electricity ^[19]. These are modeled projections for all US data centers. They replace the prior article's unsupported 2% to 3% current share and 6% to 8% 2030 range.

Carbon impact is not determined by IT electricity alone. Two facilities with the same load can have different operational emissions if their electricity comes from different generation mixes. The IEA projects that renewables will meet about half of global data-center demand growth through 2035, while natural gas and nuclear generation also expand ^[1]. Procurement contracts can support new generation, but a contract does not physically isolate a datacenter from the grid or guarantee identical carbon intensity in every operating hour.

Water consumption

Datacenter water accounting distinguishes water withdrawn from water consumed and separates direct facility use from indirect water used to generate electricity. Lawrence Berkeley National Laboratory estimated that US data centers directly consumed about 66 billion liters of water in 2023. Its 2024 report projected direct water consumption by hyperscale facilities at about 60 to 124 billion liters in 2028 ^[20]. Those national estimates do not imply that every site uses the same amount or cooling method.

Peer-reviewed research underscores that variation. Nuoa Lei and coauthors found that workload-level water use varied by more than 10,000 times across modeled conditions. The ranked determinants included server efficiency, grid water-consumption factors, utilization, cooling type, infrastructure efficiency, climate, inactive-server share, and server refresh cycle. The authors concluded that there is no single recipe for minimizing water use; the best combination is site- and workload-specific ^[21].

Direct liquid cooling can recirculate coolant and reduce dependence on room air, but the full water result depends on how the facility rejects heat and how its electricity is generated. OpenAI says its Abilene site uses a closed-loop system without traditional evaporative cooling towers, while CoreWeave says its closed-loop systems significantly reduce water consumption ^[12][18]. Those are operator descriptions of particular designs and should not be generalized to every AI datacenter.

Environmental impact summary

Scroll sideways for more →

How are AI datacenters powered? (Nuclear and renewables)

AI datacenters normally receive electricity through regional grids, sometimes supplemented by on-site generation and storage. Operators and customers use utility tariffs, power-purchase agreements, energy purchase agreements, renewable contracts, investments, and development partnerships. These mechanisms differ legally and physically. It is therefore misleading to summarize the trend as technology companies simply building and owning their own nuclear plants.

The IEA expects renewables to supply about half of global growth in datacenter electricity demand through 2035, supported by storage and the wider grid. It also projects additional natural-gas and nuclear generation, with the first small modular reactors in its outlook coming online around 2030 ^[1]. The actual mix at a site depends on regional generation, transmission, interconnection timing, reliability rules, and commercial contracts.

Several nuclear-related agreements illustrate the range of structures:

• Microsoft signed a 20-year power-purchase agreement with Constellation tied to the planned restart of Three Mile Island Unit 1 as the Crane Clean Energy Center. The 2024 announcement expected operation in 2028, but Constellation's current project page says the plant is expected to return about 835 MW to the grid in 2027 after regulatory approvals and restoration work. Microsoft would purchase energy to help match the electricity used by its datacenters in the PJM region ^[22][27].
• Google signed an agreement to purchase energy from multiple Kairos Power small modular reactors. Google said the first reactor was intended to come online by 2030, with additional deployments through 2035 and up to 500 MW supplied to US grids ^[23].
• Amazon announced a $500 million investment in X-energy intended to help advance more than 5 GW of US nuclear capacity by 2039. It also described an Energy Northwest project beginning with four reactors totaling 320 MW and expandable to 960 MW, expected in the early 2030s ^[24].
• Meta announced agreements with Vistra, TerraPower, and Oklo, alongside its earlier Constellation agreement, that it said could support up to 6.6 GW of new and existing nuclear energy by 2035. Meta described the projects as supplying grids that support its operations, including its Ohio AI cluster ^[25].

Each example contains future targets and dependencies. A signed contract is evidence of a commercial commitment, not evidence that the associated generation is already operating. Nuclear projects require licensing, construction, fuel, and grid integration. Renewable and storage projects have different lead times and operating profiles, so facility planning generally considers a portfolio rather than one technology alone ^{[1][22][23][24][25]}.

Where are AI datacenters located? (Geographic distribution)

Location decisions balance available power, grid connection time, land, fiber routes, water conditions, climate, permitting, tax policy, construction labor, and proximity to users. Training clusters can tolerate more geographic distance from end users than latency-sensitive inference, but both require dependable networking and operations. CoreWeave, for example, says its portfolio includes smaller sites near users for inference and larger sites for high-density training ^[12].

At the global level, the IEA estimated that the United States represented 45% of datacenter electricity consumption in 2024, China 25%, and Europe 15%. It also found that nearly half of US datacenter capacity was concentrated in five regional clusters and that half of US capacity under development was in existing large clusters. That concentration can create local grid bottlenecks even when the industry's global electricity share appears modest ^[1].

Recent named projects show both concentration and expansion. Stargate's public US site list includes Texas, New Mexico, Wisconsin, and Ohio ^[17]. Meta announced in July 2026 that its Richland Parish, Louisiana campus would expand to 5 GW of compute capacity with an investment above $50 billion ^[26]. CoreWeave reported facilities in six countries across the United States, Europe, and Canada at the end of 2025 ^[12]. These are disclosed examples, not a complete inventory of the industry.

Current state (2025-2026)

As of August 17, 2026, several evidence-backed conditions define the sector:

Capital deployment is high, but disclosures are not interchangeable. Amazon, Alphabet, Microsoft, and Meta each published very large 2026 capital-expenditure plans, while Oracle reported $55.663 billion for its fiscal year ended May 2026. The figures include different businesses and accounting treatments, so they should not be presented as a single audited AI-only total ^{[2][3][4][5][6]}.

Power and interconnection can delay projects. The IEA estimated that around one-fifth of planned datacenter projects could face delays without grid action. Microsoft expected its own compute capacity to remain constrained through at least 2026. These sources support power and infrastructure as important constraints without proving that power has permanently replaced chips as the sole bottleneck ^[1][4].

Cooling is becoming more diverse, not universally liquid. Very dense rack-scale systems such as GB300 NVL72 use direct liquid cooling, while current DGX B300 deployment guidance includes air-cooled patterns at 30 kW average and rear-door-assisted air cooling at 58 kW average per rack. Facility design is following platform heat load rather than one blanket transition date ^[8][9].

Training and inference both matter. CoreWeave describes a portfolio that includes large high-density training sites and smaller inference sites placed nearer users. SpaceXAI describes Colossus as serving training, fine-tuning, inference, and high-performance computing ^[12][14]. This mix affects where capacity is built, how networks are designed, and how operators schedule jobs.

Announced scale requires status labels. OpenAI reported more than 10 GW secured, Meta announced a 5 GW Louisiana campus, and CoreWeave reported 3.1 GW contracted while only more than 850 MW was active at the end of 2025 ^[12][18][26]. Reporting these numbers without "secured," "announced," "contracted," or "active" can overstate present operation.

Environmental performance is site-specific. National projections show rising electricity and water demand, while academic work shows enormous variation among workloads and locations. Accurate comparison needs operating data, boundaries, and local grid and water conditions, not only chip count or the word "liquid" ^[19][20][21].

ELI5 (Explain Like I'm 5)

An AI datacenter is a building where many powerful computer chips work together on AI jobs. The chips need fast connections so they can share information, electricity so they can run, and cooling so they do not overheat. Some racks are cooled mostly with air, some use liquid close to the chips, and some use both. A small setup can fit inside an existing datacenter, while a very large campus can need as much power planning as a town or factory. Companies often buy electricity through the grid and sign contracts that help utilities and developers build more power; that is different from every AI company owning its own power plant.

See also

• Data center
• GPU
• AI infrastructure
• AI energy consumption
• Large language model

References

1. ^International Energy Agency. "Energy and AI: Executive summary." 2025. iea.org/...executive-summary
2. ^Amazon.com, Inc. "Amazon.com Announces Fourth Quarter Results." February 5, 2026. ir.aboutamazon.com/...default
3. ^Alphabet Inc. "2025 Q4 Earnings Call." February 4, 2026. abc.xyz/...default
4. ^Microsoft Corporation. "Fiscal Year 2026 Third Quarter Earnings Conference Call." April 29, 2026. microsoft.com/...earnings-fy-2026-q3
5. ^Meta Platforms, Inc. "Meta Reports Second Quarter 2026 Results." July 29, 2026. investor.atmeta.com/...default
6. ^Oracle Corporation. "Annual Report on Form 10-K for the fiscal year ended May 31, 2026." Filed June 22, 2026. sec.gov/...orcl-20260531
7. ^NVIDIA. "Introduction to NVIDIA DGX B300 Systems." DGX B300 User Guide. Accessed August 17, 2026. docs.nvidia.com/...introduction-to-dgxb300
8. ^Hambruch, Steven, and Dennis O'Brien. "Data Center Best Practices with DGX B300." NVIDIA, version 1.0, February 27, 2026. docs.nvidia.com/...-practices-with-dgx-b300-v1.pdf
9. ^NVIDIA. "System Hardware and Components." NVIDIA GB300 NVL72 Enterprise Reference Architecture. Accessed August 17, 2026. docs.nvidia.com/...components
10. ^NVIDIA. "NVIDIA Quantum-X800 InfiniBand Platform." Accessed August 17, 2026. nvidia.com/...quantum-x800
11. ^Ultra Ethernet Consortium. "UEC Launches Specification 1.0 Transforming Ethernet for AI and HPC at Scale." June 11, 2025. ultraethernet.org/...ernet-for-ai-and-hpc-at-scale
12. ^CoreWeave, Inc. "Annual Report on Form 10-K for the year ended December 31, 2025." 2026. sec.gov/...crwv-20251231
13. ^CoreWeave, Inc. "CoreWeave and Meta Announce $21 Billion Expanded AI Infrastructure Agreement." SEC Exhibit 99.1, April 9, 2026. sec.gov/...ex991
14. ^SpaceXAI. "New Compute Partnership with Anthropic." May 6, 2026. x.ai/...anthropic-compute-partnership
15. ^SpaceXAI. "Colossus: Our gigafactory of compute." Accessed August 17, 2026. x.ai/colossus
16. ^OpenAI and SoftBank. "Announcing The Stargate Project." January 21, 2025. openai.com/...announcing-the-stargate-project
17. ^OpenAI. "OpenAI, Oracle, and SoftBank expand Stargate with five new AI data center sites." September 23, 2025, updated October 22, 2025. openai.com/...five-new-stargate-sites
18. ^OpenAI. "Building the compute infrastructure for the Intelligence Age." April 29, 2026. openai.com/...rastructure-for-the-intelligence-age
19. ^Smith, Sarah Josephine, et al. "United States Data Center Energy Usage Report: 2025 Update." Lawrence Berkeley National Laboratory, June 2026. eta-publications.lbl.gov/...ata-center-energy-2025
20. ^Shehabi, Arman, et al. "2024 United States Data Center Energy Usage Report." Lawrence Berkeley National Laboratory, December 2024. escholarship.org/...32d6m0d1
21. ^Lei, Nuoa, Jun Lu, Arman Shehabi, and Eric R. Masanet. "The water use of data center workloads: A review and assessment of key determinants." Resources, Conservation and Recycling 219 (2025): 108310. doi.org/...j.resconrec.2025.108310
22. ^Constellation Energy. "Constellation to Launch Crane Clean Energy Center, Restoring Jobs and Carbon-Free Power to the Grid." September 20, 2024. constellationenergy.com/...-Free-Power-to-The-Grid
23. ^Google. "New nuclear clean energy agreement with Kairos Power." October 14, 2024. blog.google/...iros-power-nuclear-energy-agreement
24. ^Amazon. "Amazon continues to be one of the world's leading corporate purchasers of carbon-free energy." Accessed August 17, 2026. aboutamazon.com/...te-carbon-free-energy-purchaser
25. ^Meta. "Meta Announces Nuclear Energy Projects, Unlocking Up to 6.6 GW to Power American Leadership in AI Innovation." January 2026. about.fb.com/...jects-power-american-ai-leadership
26. ^Meta. "Teachers and Local Businesses Win as Meta Expands Louisiana Data Center." July 13, 2026. about.fb.com/...meta-expands-louisiana-data-center
27. ^Constellation Energy. "Crane Clean Energy Center." Accessed August 17, 2026. constellationenergy.com/...crane-clean-energy-center